Nicolò Valigi | articles | talks

Phoenix WebSockets from the ground up

Inspired by the Phoenix is not your application presentation, I dug into Phoenix's internals to investigate how to use a custom WebSockets protocol while retaining the well-thought abstractions and patterns. I'm especially referring to Phoenix's solid pubsub backends, integration with the Cowboy HTTP server, and configuration/logging handling.

I was surprised to see that the code involved is very small and readable, with the exception of some magic patterns to shuffle between various shapes of return values, GenServer state changes, and the like. On the bright side, such code would have been hair-pulling to implement in a language without pattern matching.

What follows is a log of my journey through the code base, and all the moving pieces that come together for Phoenix's awesome WebSockets support.

Configuration

First, we take a top down look at how a Phoenix application is configured to handle WebSockets, starting from the user-provided entry point and reaching the configuration of the Cowboy HTTP server.

Endpoint (phoenix/endpoint.ex)

Your application's endpoint will use Phoenix.Endpoint and have one or more calls to its socket macro to set up the URLs that will be handling websockets. The socket macro accumulates socket information in the @phoenix_sockets attribute:

defmacro socket(path, module) do
    quote do
        @phoenix_sockets {unquote(path), unquote(module)}
    end
end

and the __before_compile__ macro creates a function that returns that attribute itself:

defmacro __before_compile__(env) do
    sockets = Module.get_attribute(env.module, :phoenix_sockets)

    quote do
        def __sockets__, do: unquote(sockets)
    end
end

The Endpoint also starts the Adapter supervision loop:

def start_link do
    Adapter.start_link(@otp_app, __MODULE__)
end

Adapter (phoenix/endpoint/adapter.ex)

The supervisor tree includes a Phoenix.Endpoint.Server supervisor, that we're going to look at next.

Endpoint.Server (phoenix/endpoint/server.ex)

Calls back to the CowboyHandler and starts supervising the child_spec it returns.

Endpoint.CowboyHandler

The child_spec function looks at the sockets defined in the Endpoint. It looks up the corresponding transport for each one and passes the bunch to the Plug.Adapters.Cowboy Cowboy adapter.

Communication

This section traces the flow of data arriving as a WebSockets message as it bubbles up towards the application logic the developer has implemented. This will take a bottom-up perspective, from the Cowboy handler to the user- provided callbacks.

Endpoint.CowboyWebSocket

When incoming messages arrive on the wire, Cowboy calls the appropriate functions on the CowboyWebSocket module. In turn, these pass the payload up to the corresponding function in the Transports.WebSocket module, such as:

def websocket_handle({opcode = :text, payload}, req, {handler, state}) do
    handle_reply req, handler, handler.ws_handle(opcode, payload, state)
end

Transport.WebSocket

Transport.WebSocket implements the Socket.Transport behaviour and uses many convenience functions defined there to handle connection, disconnection, serialization, and deserialization. In a sense, it acts as a middleman between the socket and the channel, shuffling messages back and forth when appropriate.

On connection, Transports.WebSocket calls back to the connect function in Socket.Transport. After setting up a new Socket struct that will be kept as part of the process state, that function calls back the connect function in the user-defined module to handle authentication.

Incoming messages (ws_handle), are dispatched to the right channel using the Socket.Transport.dispatch function, that figures out the right channel by looking at the HashDict:

def dispatch(%Message{} = msg, channels, socket) do
    channels
    |> HashDict.get(msg.topic)
    |> do_dispatch(msg, socket)
end

and then actually sends the payload to the Channel process:

defp do_dispatch(channel_pid, msg, _socket) do
    send(channel_pid, msg)
    :noreply
end

Especially interesting is the dispatch function that handles new topic topic subscriptions, with signature:

defp do_dispatch(nil, %{event: "phx_join", topic: topic} = msg, socket) do

that checks whether the user has defined a channel corresponding to the current topic,

if channel = socket.handler.__channel__(topic, socket.transport_name) do
    socket = %Socket{socket | topic: topic, channel: channel}

has the Socket join it,

Phoenix.Channel.Server.join(socket, msg.payload)

and replies positively to the client:

{:joined, pid, %Reply{ref: msg.ref, topic: topic, status: :ok, payload: response}}

Channel.Server

As the Channel module itself is only a behaviour that user code needs to adhere to, most of the actual functionality is implemented in the Genserver at Channel.Server. Each new socket connection causes a new Channel.Server to be spun up. Precisely, it's the transport that calls the join function in the Channel.Server:

def join(socket, auth_payload) do
    ref = make_ref()
    case GenServer.start_link(__MODULE__, {socket, auth_payload, self(), ref}) do

in the init GenServer callback, the channel asks the Pubsub.Server to subscribe to its topic:

PubSub.subscribe(socket.pubsub_server, self(), socket.topic,

and then messages itself back the result of the subscription, so that it can handle messages.

Broadcasts are sent through the pubsub server:

def broadcast(pubsub_server, topic, event, payload) do
    PubSub.broadcast pubsub_server, topic, %Broadcast{
      topic: topic,
      event: event,
      payload: payload
    }
end

Finally, messages received from the pubsub server are handled by the user's application logic implemented in the channel's handle_in function:

def handle_info(%Message{topic: topic, event: event, payload: payload, ref: ref},
                %{topic: topic} = socket) do
    event
    |> socket.channel.handle_in(payload, put_in(socket.ref, ref))
    |> handle_result(:handle_in)
end

Conclusion

While well thought-out, the internals are pretty tightly coupled to the topic- based nature of Phoenix's WS protocol. Transports, channels, and the pubsub server all operate by switching messages based on their topic. While flexibile enough for most clean-slate real time applications, this choice makes it somewhat hard to retro-fit an existing protocol without overhauling significant parts of the stack.